On August 14 I will give there a plenary seminar (sort of a keynote lecture), which I decided to title "The Second AI Revolution in Fundamental Physics". If you have not been hybernating or brain-dead in the past 15 years, you will probably know that an AI revolution already took place in fundamental physics research: from 2012 onwards the way physicists have been extracting information and inference from the complex data produced by their experiments has been through machine learning methods and increasingly complex deep learning models. But what is the second revolution I am talking about, then?
(Above, the web page of the Swedish Physics Days)
A revolution in the making
Well, that's the one that is slowly getting prepared, by a small group of pioneers that has grown a legion during the past 4-5 years. In 2019 I founded the MODE collaboration (https://mode-collaboration.github.io), a group that includes physicists and computer scientists who are convinced of the transformative potential of AI methods for the design of scientific instruments and experiments - the hardware that elicits data from subnuclear processes, that is.
This is a different paradigm change from the one that happened in 2012, and one much harder to bring about. Since 2020, we have been working at demonstrating in easy use cases how the use of deep learning and complete models of experiments (parametrizing both the hardware and the software) is a powerful enabler of holistic optimization, bringing forth huge potential benefits and effectiveness of our instruments.
Then, two years ago the EUCAIF coalition (https://eucaif.org) was formed. Broader in scope and targeting not only experiment optimization but a wealth of applications of AI to fundamental physics, EUCAIF has organized two large conferences (2024 in Amsterdam, and 2025 in Cagliari), gathering the most active researchers who embrace AI for their studies. It has been wonderful to see how much progress we have been making with deep learning models, and especially to observe how physicists are active developers of innovative machine learning techniques, as our problems demand custom solutions that are not always available.
It has been difficult to work in almost complete isolation at a few optimization problems in these past few years, with little support from funding agencies (with the notable exception of Iris-HEP and a few others) and the constant rejection of our proposals from EU programs (MSCA, EIC, ERC). The reviewers in those panels failed to realize the importance of integrating AI in the design of hardware for future experiments until now.
But I sense that the wind is changing.
So, a second revolution is in the making in fundamental physics. Outside research, in the profit-driven world, we have in recent years seen the advent of true AI with ChatGPT and other large language models revolutionizing the market. Those innovations are still waiting to be integrated in the workflow of hard science because we have not had the money to instantiate them yet. But with a lot of effort and good will, this is happening now, and it will change everything.
If you are a detector builder reading these lines, your feelings might be a mix of scepticism and worry. Are they coming for your job, too? "Nah, designing a particle detector is something too complex for a machine". Well, you are both right and wrong with the scepticism. We are not close to study the parameter space of a large collider experiment yet, but we are already showing significant gains from holistic optimization of medium-size instruments. On the matter of being worried, though, you are completely wrong - you are looking at the question the wrong way.
The software pipelines we have been developing are not meant to substitute humans in taking critical decisions on how to design complex experiments. They are meant to assist humans doing that! By scanning the large parameter space of experiment hardware and software, the machine can discover new advantageous ways to use our technology for our measurements, but the machine cannot include in its modeling some factors that remain under human control and that cannot be inserted in a software program or utility function. I will make a simple example to clarify why humans will remain in the driver's seat in this task.
Why detector builders should not be scared
Large experiments like CMS and ATLAS are built by consortia of many institutes from tens of different contributing nations. Each of these has interest in developing specific technologies, so they lobby their funding agencies to get money for their hardware R&D. These efforts end up contributing to the final instrument that is approved and constructed. This is made up of tens of different subsystems, each of which has its own budget, its own driving technology, and specific funding backing it. An AI system modeling the full detector might in principle be capable of exploring completely different solutions for the overall design, and propose wild rearrangements of the budget allocation to each subsystem. But those suggestions would not sit well with the consortium. You can't just tell the calorimeter guys to give half of their budget to the tracking guys "because this will improve the overall scientific output of the experiment by 20%".
The correct way of thinking at the matter is of an AI system that proposes alternative layouts that have different cost redistributions and performances, with estimated overall scientific outputs. Those layouts would then be considered by the detector designers, and their parameters could then be continuously varied to sample configurations still close to the Pareto front while more acceptable from the internal politics of the experiments. The true innovation here - or the meta-innovation to be precise - is not the identification of a parameter configuration that guarantees a large sensitivity improvement at same global cost. There will be many of these instances, for sure; I call these "AI-discovered layouts".
A meta-innovation
The true meta-innovation is instead the fact that with a differentiable model of the whole experiment you can then study the vicinity of those solutions, in a continuous way. This was never doable with discrete scans (which are all you can do when your simulations are not differentiable). That, I think, will be the true revolutionary bit. Once we empower ourselves with these formidable helpers, we will be able to choose the design of our experiments with a full knowledge of what is possible with the available technology, without having to rely on well-grounded construction paradigms that may completely miss the large gains coming from exploitation of subtle interplay of different sub-detector configurations.
There would be books to write about this topic, and indeed I published a dozen articles on this matter in the past few years. If you are interested, drop me a line!
Comments